Answers are grounded in this paper and cite their sources.
Accurate chemical kinetics are crucial for reacting-flow simulations but detailed mechanisms remain computational prohibitive. To fill this gap, we formulate a self-evolving neural network (SENN) methodology with a species–reaction–species topology, where thermochemical states feed reaction neurons whose outputs represent the collective net species production rates. The SENN undergoes episodic training in which the network is exposed to a subset of data and gradient-based optimization is coupled with stabilized Hebbian plasticity to strengthen persistently co-active connections. After each episode, a stoichiometry-graph guided pruning step removes neurons and edges with consistently low influence. The resulting neural network (NN) constitutes a compact, chemistry-consistent reduced mechanism that faithfully predicts the detailed kinetics over a broad range of thermochemical states. The SENN methodology is demonstrated for H 2 /air combustion using a 21-reaction model as the baseline detailed kinetics. First, we demonstrate that the reduced kinetic models from SENN accurately predict net species production rate with negligible mean-squared error (MSE), closely capture ignition delay and flame speed over a wide range of thermochemical states, and provide up to 4.5x computational speedup. Second, we develop a SENN-centered adaptive modeling framework that dynamically adapts the reduced mechanisms on the fly to accurately capture the reaction characteristics while efficiently resolving slower chemical regimes, reproducing the temperature and heat release rate evolution with good agreement with the detailed kinetics and achieving an additional computational speedup of approximately 6 . 73 × . The SENN methodology offers a new direction in developing an interpretable and scalable scientific machine learning approach for chemical kinetic reduction that balances computational efficiency and modeling accuracy in complex reacting flow simulations. Novelty and significance statement: This work introduces a self-evolving neural network (SENN) methodology that autonomously evolves its structure to learn a sparse, chemistry-consistent kinetic model. SENN features three key innovations: (1) a species–reaction–species topology enabling adaptive structure evolution; (2) episodic training, which reduces training cost and guides physics-informed adaptation; and (3) stoichiometry-guided graph pruning reinforced by Hebbian consolidation and gradient saliency, yielding reduced kinetic mechanisms that preserve key combustion observables with substantially lower inference cost. SENN provides a generalizable framework for extracting skeletal mechanisms, quantifying reaction-pathway importance, and managing the sparsity–accuracy tradeoff. The approach enables robust mechanism reduction for practical combustion systems, offers interpretable diagnostics of dominant reactions, and paves the way for extensions to complex fuels and broader thermochemical regimes.
No takes yet. Share an insight, caveat, or question.
Nasir et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: